1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2016 Intel Corporation
7 #include <linux/virtio_net.h>
10 #include <rte_memcpy.h>
12 #include <rte_ether.h>
14 #include <rte_vhost.h>
19 #include <rte_spinlock.h>
20 #include <rte_malloc.h>
21 #include <rte_vhost_async.h>
26 #define MAX_BATCH_LEN 256
28 static __rte_always_inline bool
29 rxvq_is_mergeable(struct virtio_net *dev)
31 return dev->features & (1ULL << VIRTIO_NET_F_MRG_RXBUF);
34 static __rte_always_inline bool
35 virtio_net_is_inorder(struct virtio_net *dev)
37 return dev->features & (1ULL << VIRTIO_F_IN_ORDER);
41 is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
43 return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
47 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
49 struct batch_copy_elem *elem = vq->batch_copy_elems;
50 uint16_t count = vq->batch_copy_nb_elems;
53 for (i = 0; i < count; i++) {
54 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
55 vhost_log_cache_write_iova(dev, vq, elem[i].log_addr,
57 PRINT_PACKET(dev, (uintptr_t)elem[i].dst, elem[i].len, 0);
60 vq->batch_copy_nb_elems = 0;
64 do_data_copy_dequeue(struct vhost_virtqueue *vq)
66 struct batch_copy_elem *elem = vq->batch_copy_elems;
67 uint16_t count = vq->batch_copy_nb_elems;
70 for (i = 0; i < count; i++)
71 rte_memcpy(elem[i].dst, elem[i].src, elem[i].len);
73 vq->batch_copy_nb_elems = 0;
76 static __rte_always_inline void
77 do_flush_shadow_used_ring_split(struct virtio_net *dev,
78 struct vhost_virtqueue *vq,
79 uint16_t to, uint16_t from, uint16_t size)
81 rte_memcpy(&vq->used->ring[to],
82 &vq->shadow_used_split[from],
83 size * sizeof(struct vring_used_elem));
84 vhost_log_cache_used_vring(dev, vq,
85 offsetof(struct vring_used, ring[to]),
86 size * sizeof(struct vring_used_elem));
89 static __rte_always_inline void
90 flush_shadow_used_ring_split(struct virtio_net *dev, struct vhost_virtqueue *vq)
92 uint16_t used_idx = vq->last_used_idx & (vq->size - 1);
94 if (used_idx + vq->shadow_used_idx <= vq->size) {
95 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0,
100 /* update used ring interval [used_idx, vq->size] */
101 size = vq->size - used_idx;
102 do_flush_shadow_used_ring_split(dev, vq, used_idx, 0, size);
104 /* update the left half used ring interval [0, left_size] */
105 do_flush_shadow_used_ring_split(dev, vq, 0, size,
106 vq->shadow_used_idx - size);
108 vq->last_used_idx += vq->shadow_used_idx;
110 vhost_log_cache_sync(dev, vq);
112 __atomic_add_fetch(&vq->used->idx, vq->shadow_used_idx,
114 vq->shadow_used_idx = 0;
115 vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
116 sizeof(vq->used->idx));
119 static __rte_always_inline void
120 update_shadow_used_ring_split(struct vhost_virtqueue *vq,
121 uint16_t desc_idx, uint32_t len)
123 uint16_t i = vq->shadow_used_idx++;
125 vq->shadow_used_split[i].id = desc_idx;
126 vq->shadow_used_split[i].len = len;
129 static __rte_always_inline void
130 vhost_flush_enqueue_shadow_packed(struct virtio_net *dev,
131 struct vhost_virtqueue *vq)
134 uint16_t used_idx = vq->last_used_idx;
135 uint16_t head_idx = vq->last_used_idx;
136 uint16_t head_flags = 0;
138 /* Split loop in two to save memory barriers */
139 for (i = 0; i < vq->shadow_used_idx; i++) {
140 vq->desc_packed[used_idx].id = vq->shadow_used_packed[i].id;
141 vq->desc_packed[used_idx].len = vq->shadow_used_packed[i].len;
143 used_idx += vq->shadow_used_packed[i].count;
144 if (used_idx >= vq->size)
145 used_idx -= vq->size;
148 /* The ordering for storing desc flags needs to be enforced. */
149 rte_atomic_thread_fence(__ATOMIC_RELEASE);
151 for (i = 0; i < vq->shadow_used_idx; i++) {
154 if (vq->shadow_used_packed[i].len)
155 flags = VRING_DESC_F_WRITE;
159 if (vq->used_wrap_counter) {
160 flags |= VRING_DESC_F_USED;
161 flags |= VRING_DESC_F_AVAIL;
163 flags &= ~VRING_DESC_F_USED;
164 flags &= ~VRING_DESC_F_AVAIL;
168 vq->desc_packed[vq->last_used_idx].flags = flags;
170 vhost_log_cache_used_vring(dev, vq,
172 sizeof(struct vring_packed_desc),
173 sizeof(struct vring_packed_desc));
175 head_idx = vq->last_used_idx;
179 vq_inc_last_used_packed(vq, vq->shadow_used_packed[i].count);
182 vq->desc_packed[head_idx].flags = head_flags;
184 vhost_log_cache_used_vring(dev, vq,
186 sizeof(struct vring_packed_desc),
187 sizeof(struct vring_packed_desc));
189 vq->shadow_used_idx = 0;
190 vhost_log_cache_sync(dev, vq);
193 static __rte_always_inline void
194 vhost_flush_dequeue_shadow_packed(struct virtio_net *dev,
195 struct vhost_virtqueue *vq)
197 struct vring_used_elem_packed *used_elem = &vq->shadow_used_packed[0];
199 vq->desc_packed[vq->shadow_last_used_idx].id = used_elem->id;
200 /* desc flags is the synchronization point for virtio packed vring */
201 __atomic_store_n(&vq->desc_packed[vq->shadow_last_used_idx].flags,
202 used_elem->flags, __ATOMIC_RELEASE);
204 vhost_log_cache_used_vring(dev, vq, vq->shadow_last_used_idx *
205 sizeof(struct vring_packed_desc),
206 sizeof(struct vring_packed_desc));
207 vq->shadow_used_idx = 0;
208 vhost_log_cache_sync(dev, vq);
211 static __rte_always_inline void
212 vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
213 struct vhost_virtqueue *vq,
219 uint16_t last_used_idx;
220 struct vring_packed_desc *desc_base;
222 last_used_idx = vq->last_used_idx;
223 desc_base = &vq->desc_packed[last_used_idx];
225 flags = PACKED_DESC_ENQUEUE_USED_FLAG(vq->used_wrap_counter);
227 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
228 desc_base[i].id = ids[i];
229 desc_base[i].len = lens[i];
232 rte_atomic_thread_fence(__ATOMIC_RELEASE);
234 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
235 desc_base[i].flags = flags;
238 vhost_log_cache_used_vring(dev, vq, last_used_idx *
239 sizeof(struct vring_packed_desc),
240 sizeof(struct vring_packed_desc) *
242 vhost_log_cache_sync(dev, vq);
244 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
247 static __rte_always_inline void
248 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
251 vq->shadow_used_packed[0].id = id;
253 if (!vq->shadow_used_idx) {
254 vq->shadow_last_used_idx = vq->last_used_idx;
255 vq->shadow_used_packed[0].flags =
256 PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
257 vq->shadow_used_packed[0].len = 0;
258 vq->shadow_used_packed[0].count = 1;
259 vq->shadow_used_idx++;
262 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
265 static __rte_always_inline void
266 vhost_shadow_dequeue_batch_packed(struct virtio_net *dev,
267 struct vhost_virtqueue *vq,
274 flags = PACKED_DESC_DEQUEUE_USED_FLAG(vq->used_wrap_counter);
276 if (!vq->shadow_used_idx) {
277 vq->shadow_last_used_idx = vq->last_used_idx;
278 vq->shadow_used_packed[0].id = ids[0];
279 vq->shadow_used_packed[0].len = 0;
280 vq->shadow_used_packed[0].count = 1;
281 vq->shadow_used_packed[0].flags = flags;
282 vq->shadow_used_idx++;
287 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE) {
288 vq->desc_packed[vq->last_used_idx + i].id = ids[i];
289 vq->desc_packed[vq->last_used_idx + i].len = 0;
292 rte_atomic_thread_fence(__ATOMIC_RELEASE);
293 vhost_for_each_try_unroll(i, begin, PACKED_BATCH_SIZE)
294 vq->desc_packed[vq->last_used_idx + i].flags = flags;
296 vhost_log_cache_used_vring(dev, vq, vq->last_used_idx *
297 sizeof(struct vring_packed_desc),
298 sizeof(struct vring_packed_desc) *
300 vhost_log_cache_sync(dev, vq);
302 vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
305 static __rte_always_inline void
306 vhost_shadow_dequeue_single_packed(struct vhost_virtqueue *vq,
312 flags = vq->desc_packed[vq->last_used_idx].flags;
313 if (vq->used_wrap_counter) {
314 flags |= VRING_DESC_F_USED;
315 flags |= VRING_DESC_F_AVAIL;
317 flags &= ~VRING_DESC_F_USED;
318 flags &= ~VRING_DESC_F_AVAIL;
321 if (!vq->shadow_used_idx) {
322 vq->shadow_last_used_idx = vq->last_used_idx;
324 vq->shadow_used_packed[0].id = buf_id;
325 vq->shadow_used_packed[0].len = 0;
326 vq->shadow_used_packed[0].flags = flags;
327 vq->shadow_used_idx++;
329 vq->desc_packed[vq->last_used_idx].id = buf_id;
330 vq->desc_packed[vq->last_used_idx].len = 0;
331 vq->desc_packed[vq->last_used_idx].flags = flags;
334 vq_inc_last_used_packed(vq, count);
337 static __rte_always_inline void
338 vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
344 vq->shadow_used_packed[0].id = buf_id;
346 flags = vq->desc_packed[vq->last_used_idx].flags;
347 if (vq->used_wrap_counter) {
348 flags |= VRING_DESC_F_USED;
349 flags |= VRING_DESC_F_AVAIL;
351 flags &= ~VRING_DESC_F_USED;
352 flags &= ~VRING_DESC_F_AVAIL;
355 if (!vq->shadow_used_idx) {
356 vq->shadow_last_used_idx = vq->last_used_idx;
357 vq->shadow_used_packed[0].len = 0;
358 vq->shadow_used_packed[0].flags = flags;
359 vq->shadow_used_idx++;
362 vq_inc_last_used_packed(vq, count);
365 static __rte_always_inline void
366 vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
370 uint16_t num_buffers)
374 for (i = 0; i < num_buffers; i++) {
375 /* enqueue shadow flush action aligned with batch num */
376 if (!vq->shadow_used_idx)
377 vq->shadow_aligned_idx = vq->last_used_idx &
379 vq->shadow_used_packed[vq->shadow_used_idx].id = id[i];
380 vq->shadow_used_packed[vq->shadow_used_idx].len = len[i];
381 vq->shadow_used_packed[vq->shadow_used_idx].count = count[i];
382 vq->shadow_aligned_idx += count[i];
383 vq->shadow_used_idx++;
387 static __rte_always_inline void
388 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
389 struct vhost_virtqueue *vq,
393 uint16_t num_buffers)
395 vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
397 if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
398 do_data_copy_enqueue(dev, vq);
399 vhost_flush_enqueue_shadow_packed(dev, vq);
403 /* avoid write operation when necessary, to lessen cache issues */
404 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
405 if ((var) != (val)) \
409 static __rte_always_inline void
410 virtio_enqueue_offload(struct rte_mbuf *m_buf, struct virtio_net_hdr *net_hdr)
412 uint64_t csum_l4 = m_buf->ol_flags & RTE_MBUF_F_TX_L4_MASK;
414 if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
415 csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
418 net_hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
419 net_hdr->csum_start = m_buf->l2_len + m_buf->l3_len;
422 case RTE_MBUF_F_TX_TCP_CKSUM:
423 net_hdr->csum_offset = (offsetof(struct rte_tcp_hdr,
426 case RTE_MBUF_F_TX_UDP_CKSUM:
427 net_hdr->csum_offset = (offsetof(struct rte_udp_hdr,
430 case RTE_MBUF_F_TX_SCTP_CKSUM:
431 net_hdr->csum_offset = (offsetof(struct rte_sctp_hdr,
436 ASSIGN_UNLESS_EQUAL(net_hdr->csum_start, 0);
437 ASSIGN_UNLESS_EQUAL(net_hdr->csum_offset, 0);
438 ASSIGN_UNLESS_EQUAL(net_hdr->flags, 0);
441 /* IP cksum verification cannot be bypassed, then calculate here */
442 if (m_buf->ol_flags & RTE_MBUF_F_TX_IP_CKSUM) {
443 struct rte_ipv4_hdr *ipv4_hdr;
445 ipv4_hdr = rte_pktmbuf_mtod_offset(m_buf, struct rte_ipv4_hdr *,
447 ipv4_hdr->hdr_checksum = 0;
448 ipv4_hdr->hdr_checksum = rte_ipv4_cksum(ipv4_hdr);
451 if (m_buf->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
452 if (m_buf->ol_flags & RTE_MBUF_F_TX_IPV4)
453 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
455 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
456 net_hdr->gso_size = m_buf->tso_segsz;
457 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len
459 } else if (m_buf->ol_flags & RTE_MBUF_F_TX_UDP_SEG) {
460 net_hdr->gso_type = VIRTIO_NET_HDR_GSO_UDP;
461 net_hdr->gso_size = m_buf->tso_segsz;
462 net_hdr->hdr_len = m_buf->l2_len + m_buf->l3_len +
465 ASSIGN_UNLESS_EQUAL(net_hdr->gso_type, 0);
466 ASSIGN_UNLESS_EQUAL(net_hdr->gso_size, 0);
467 ASSIGN_UNLESS_EQUAL(net_hdr->hdr_len, 0);
471 static __rte_always_inline int
472 map_one_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
473 struct buf_vector *buf_vec, uint16_t *vec_idx,
474 uint64_t desc_iova, uint64_t desc_len, uint8_t perm)
476 uint16_t vec_id = *vec_idx;
480 uint64_t desc_chunck_len = desc_len;
482 if (unlikely(vec_id >= BUF_VECTOR_MAX))
485 desc_addr = vhost_iova_to_vva(dev, vq,
489 if (unlikely(!desc_addr))
492 rte_prefetch0((void *)(uintptr_t)desc_addr);
494 buf_vec[vec_id].buf_iova = desc_iova;
495 buf_vec[vec_id].buf_addr = desc_addr;
496 buf_vec[vec_id].buf_len = desc_chunck_len;
498 desc_len -= desc_chunck_len;
499 desc_iova += desc_chunck_len;
507 static __rte_always_inline int
508 fill_vec_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
509 uint32_t avail_idx, uint16_t *vec_idx,
510 struct buf_vector *buf_vec, uint16_t *desc_chain_head,
511 uint32_t *desc_chain_len, uint8_t perm)
513 uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
514 uint16_t vec_id = *vec_idx;
517 uint32_t nr_descs = vq->size;
519 struct vring_desc *descs = vq->desc;
520 struct vring_desc *idesc = NULL;
522 if (unlikely(idx >= vq->size))
525 *desc_chain_head = idx;
527 if (vq->desc[idx].flags & VRING_DESC_F_INDIRECT) {
528 dlen = vq->desc[idx].len;
529 nr_descs = dlen / sizeof(struct vring_desc);
530 if (unlikely(nr_descs > vq->size))
533 descs = (struct vring_desc *)(uintptr_t)
534 vhost_iova_to_vva(dev, vq, vq->desc[idx].addr,
537 if (unlikely(!descs))
540 if (unlikely(dlen < vq->desc[idx].len)) {
542 * The indirect desc table is not contiguous
543 * in process VA space, we have to copy it.
545 idesc = vhost_alloc_copy_ind_table(dev, vq,
546 vq->desc[idx].addr, vq->desc[idx].len);
547 if (unlikely(!idesc))
557 if (unlikely(idx >= nr_descs || cnt++ >= nr_descs)) {
558 free_ind_table(idesc);
562 dlen = descs[idx].len;
565 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
566 descs[idx].addr, dlen,
568 free_ind_table(idesc);
572 if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
575 idx = descs[idx].next;
578 *desc_chain_len = len;
581 if (unlikely(!!idesc))
582 free_ind_table(idesc);
588 * Returns -1 on fail, 0 on success
591 reserve_avail_buf_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
592 uint32_t size, struct buf_vector *buf_vec,
593 uint16_t *num_buffers, uint16_t avail_head,
597 uint16_t vec_idx = 0;
598 uint16_t max_tries, tries = 0;
600 uint16_t head_idx = 0;
604 cur_idx = vq->last_avail_idx;
606 if (rxvq_is_mergeable(dev))
607 max_tries = vq->size - 1;
612 if (unlikely(cur_idx == avail_head))
615 * if we tried all available ring items, and still
616 * can't get enough buf, it means something abnormal
619 if (unlikely(++tries > max_tries))
622 if (unlikely(fill_vec_buf_split(dev, vq, cur_idx,
625 VHOST_ACCESS_RW) < 0))
627 len = RTE_MIN(len, size);
628 update_shadow_used_ring_split(vq, head_idx, len);
640 static __rte_always_inline int
641 fill_vec_buf_packed_indirect(struct virtio_net *dev,
642 struct vhost_virtqueue *vq,
643 struct vring_packed_desc *desc, uint16_t *vec_idx,
644 struct buf_vector *buf_vec, uint32_t *len, uint8_t perm)
648 uint16_t vec_id = *vec_idx;
650 struct vring_packed_desc *descs, *idescs = NULL;
653 descs = (struct vring_packed_desc *)(uintptr_t)
654 vhost_iova_to_vva(dev, vq, desc->addr, &dlen, VHOST_ACCESS_RO);
655 if (unlikely(!descs))
658 if (unlikely(dlen < desc->len)) {
660 * The indirect desc table is not contiguous
661 * in process VA space, we have to copy it.
663 idescs = vhost_alloc_copy_ind_table(dev,
664 vq, desc->addr, desc->len);
665 if (unlikely(!idescs))
671 nr_descs = desc->len / sizeof(struct vring_packed_desc);
672 if (unlikely(nr_descs >= vq->size)) {
673 free_ind_table(idescs);
677 for (i = 0; i < nr_descs; i++) {
678 if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
679 free_ind_table(idescs);
685 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
692 if (unlikely(!!idescs))
693 free_ind_table(idescs);
698 static __rte_always_inline int
699 fill_vec_buf_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
700 uint16_t avail_idx, uint16_t *desc_count,
701 struct buf_vector *buf_vec, uint16_t *vec_idx,
702 uint16_t *buf_id, uint32_t *len, uint8_t perm)
704 bool wrap_counter = vq->avail_wrap_counter;
705 struct vring_packed_desc *descs = vq->desc_packed;
706 uint16_t vec_id = *vec_idx;
709 if (avail_idx < vq->last_avail_idx)
713 * Perform a load-acquire barrier in desc_is_avail to
714 * enforce the ordering between desc flags and desc
717 if (unlikely(!desc_is_avail(&descs[avail_idx], wrap_counter)))
724 if (unlikely(vec_id >= BUF_VECTOR_MAX))
727 if (unlikely(*desc_count >= vq->size))
731 *buf_id = descs[avail_idx].id;
733 if (descs[avail_idx].flags & VRING_DESC_F_INDIRECT) {
734 if (unlikely(fill_vec_buf_packed_indirect(dev, vq,
740 dlen = descs[avail_idx].len;
743 if (unlikely(map_one_desc(dev, vq, buf_vec, &vec_id,
744 descs[avail_idx].addr,
750 if ((descs[avail_idx].flags & VRING_DESC_F_NEXT) == 0)
753 if (++avail_idx >= vq->size) {
754 avail_idx -= vq->size;
764 static __rte_noinline void
765 copy_vnet_hdr_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
766 struct buf_vector *buf_vec,
767 struct virtio_net_hdr_mrg_rxbuf *hdr)
770 uint64_t remain = dev->vhost_hlen;
771 uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
772 uint64_t iova = buf_vec->buf_iova;
775 len = RTE_MIN(remain,
777 dst = buf_vec->buf_addr;
778 rte_memcpy((void *)(uintptr_t)dst,
779 (void *)(uintptr_t)src,
782 PRINT_PACKET(dev, (uintptr_t)dst,
784 vhost_log_cache_write_iova(dev, vq,
794 static __rte_always_inline int
795 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
796 struct rte_mbuf *m, struct buf_vector *buf_vec,
797 uint16_t nr_vec, uint16_t num_buffers)
799 uint32_t vec_idx = 0;
800 uint32_t mbuf_offset, mbuf_avail;
801 uint32_t buf_offset, buf_avail;
802 uint64_t buf_addr, buf_iova, buf_len;
805 struct rte_mbuf *hdr_mbuf;
806 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
807 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
810 if (unlikely(m == NULL)) {
815 buf_addr = buf_vec[vec_idx].buf_addr;
816 buf_iova = buf_vec[vec_idx].buf_iova;
817 buf_len = buf_vec[vec_idx].buf_len;
819 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
826 if (unlikely(buf_len < dev->vhost_hlen)) {
827 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
830 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
832 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
833 dev->vid, num_buffers);
835 if (unlikely(buf_len < dev->vhost_hlen)) {
836 buf_offset = dev->vhost_hlen - buf_len;
838 buf_addr = buf_vec[vec_idx].buf_addr;
839 buf_iova = buf_vec[vec_idx].buf_iova;
840 buf_len = buf_vec[vec_idx].buf_len;
841 buf_avail = buf_len - buf_offset;
843 buf_offset = dev->vhost_hlen;
844 buf_avail = buf_len - dev->vhost_hlen;
847 mbuf_avail = rte_pktmbuf_data_len(m);
849 while (mbuf_avail != 0 || m->next != NULL) {
850 /* done with current buf, get the next one */
851 if (buf_avail == 0) {
853 if (unlikely(vec_idx >= nr_vec)) {
858 buf_addr = buf_vec[vec_idx].buf_addr;
859 buf_iova = buf_vec[vec_idx].buf_iova;
860 buf_len = buf_vec[vec_idx].buf_len;
866 /* done with current mbuf, get the next one */
867 if (mbuf_avail == 0) {
871 mbuf_avail = rte_pktmbuf_data_len(m);
875 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
876 if (rxvq_is_mergeable(dev))
877 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
880 if (unlikely(hdr == &tmp_hdr)) {
881 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
883 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
885 vhost_log_cache_write_iova(dev, vq,
893 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
895 if (likely(cpy_len > MAX_BATCH_LEN ||
896 vq->batch_copy_nb_elems >= vq->size)) {
897 rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
898 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
900 vhost_log_cache_write_iova(dev, vq,
901 buf_iova + buf_offset,
903 PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
906 batch_copy[vq->batch_copy_nb_elems].dst =
907 (void *)((uintptr_t)(buf_addr + buf_offset));
908 batch_copy[vq->batch_copy_nb_elems].src =
909 rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
910 batch_copy[vq->batch_copy_nb_elems].log_addr =
911 buf_iova + buf_offset;
912 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
913 vq->batch_copy_nb_elems++;
916 mbuf_avail -= cpy_len;
917 mbuf_offset += cpy_len;
918 buf_avail -= cpy_len;
919 buf_offset += cpy_len;
927 static __rte_always_inline int
928 async_iter_initialize(struct vhost_async *async)
930 struct rte_vhost_iov_iter *iter;
932 if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
933 VHOST_LOG_DATA(ERR, "no more async iovec available\n");
937 iter = async->iov_iter + async->iter_idx;
938 iter->iov = async->iovec + async->iovec_idx;
944 static __rte_always_inline int
945 async_iter_add_iovec(struct vhost_async *async, void *src, void *dst, size_t len)
947 struct rte_vhost_iov_iter *iter;
948 struct rte_vhost_iovec *iovec;
950 if (unlikely(async->iovec_idx >= VHOST_MAX_ASYNC_VEC)) {
951 static bool vhost_max_async_vec_log;
953 if (!vhost_max_async_vec_log) {
954 VHOST_LOG_DATA(ERR, "no more async iovec available\n");
955 vhost_max_async_vec_log = true;
961 iter = async->iov_iter + async->iter_idx;
962 iovec = async->iovec + async->iovec_idx;
964 iovec->src_addr = src;
965 iovec->dst_addr = dst;
974 static __rte_always_inline void
975 async_iter_finalize(struct vhost_async *async)
980 static __rte_always_inline void
981 async_iter_cancel(struct vhost_async *async)
983 struct rte_vhost_iov_iter *iter;
985 iter = async->iov_iter + async->iter_idx;
986 async->iovec_idx -= iter->nr_segs;
991 static __rte_always_inline void
992 async_iter_reset(struct vhost_async *async)
995 async->iovec_idx = 0;
998 static __rte_always_inline int
999 async_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
1000 struct rte_mbuf *m, struct buf_vector *buf_vec,
1001 uint16_t nr_vec, uint16_t num_buffers)
1003 struct vhost_async *async = vq->async;
1004 struct rte_mbuf *hdr_mbuf;
1005 struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
1006 uint64_t buf_addr, buf_iova;
1008 uint64_t mapped_len;
1009 uint32_t vec_idx = 0;
1010 uint32_t mbuf_offset, mbuf_avail;
1011 uint32_t buf_offset, buf_avail;
1012 uint32_t cpy_len, buf_len;
1016 if (unlikely(m == NULL))
1019 buf_addr = buf_vec[vec_idx].buf_addr;
1020 buf_iova = buf_vec[vec_idx].buf_iova;
1021 buf_len = buf_vec[vec_idx].buf_len;
1023 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1))
1027 hdr_addr = buf_addr;
1028 if (unlikely(buf_len < dev->vhost_hlen)) {
1029 memset(&tmp_hdr, 0, sizeof(struct virtio_net_hdr_mrg_rxbuf));
1032 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
1034 VHOST_LOG_DATA(DEBUG, "(%d) RX: num merge buffers %d\n",
1035 dev->vid, num_buffers);
1037 if (unlikely(buf_len < dev->vhost_hlen)) {
1038 buf_offset = dev->vhost_hlen - buf_len;
1040 buf_addr = buf_vec[vec_idx].buf_addr;
1041 buf_iova = buf_vec[vec_idx].buf_iova;
1042 buf_len = buf_vec[vec_idx].buf_len;
1043 buf_avail = buf_len - buf_offset;
1045 buf_offset = dev->vhost_hlen;
1046 buf_avail = buf_len - dev->vhost_hlen;
1049 mbuf_avail = rte_pktmbuf_data_len(m);
1052 if (async_iter_initialize(async))
1055 while (mbuf_avail != 0 || m->next != NULL) {
1056 /* done with current buf, get the next one */
1057 if (buf_avail == 0) {
1059 if (unlikely(vec_idx >= nr_vec))
1062 buf_addr = buf_vec[vec_idx].buf_addr;
1063 buf_iova = buf_vec[vec_idx].buf_iova;
1064 buf_len = buf_vec[vec_idx].buf_len;
1067 buf_avail = buf_len;
1070 /* done with current mbuf, get the next one */
1071 if (mbuf_avail == 0) {
1075 mbuf_avail = rte_pktmbuf_data_len(m);
1079 virtio_enqueue_offload(hdr_mbuf, &hdr->hdr);
1080 if (rxvq_is_mergeable(dev))
1081 ASSIGN_UNLESS_EQUAL(hdr->num_buffers,
1084 if (unlikely(hdr == &tmp_hdr)) {
1085 copy_vnet_hdr_to_desc(dev, vq, buf_vec, hdr);
1087 PRINT_PACKET(dev, (uintptr_t)hdr_addr,
1088 dev->vhost_hlen, 0);
1089 vhost_log_cache_write_iova(dev, vq,
1090 buf_vec[0].buf_iova,
1097 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
1099 while (unlikely(cpy_len)) {
1100 hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
1101 buf_iova + buf_offset,
1102 cpy_len, &mapped_len);
1103 if (unlikely(!hpa)) {
1104 VHOST_LOG_DATA(ERR, "(%d) %s: failed to get hpa.\n",
1105 dev->vid, __func__);
1109 if (unlikely(async_iter_add_iovec(async,
1110 (void *)(uintptr_t)rte_pktmbuf_iova_offset(m,
1112 hpa, (size_t)mapped_len)))
1115 cpy_len -= (uint32_t)mapped_len;
1116 mbuf_avail -= (uint32_t)mapped_len;
1117 mbuf_offset += (uint32_t)mapped_len;
1118 buf_avail -= (uint32_t)mapped_len;
1119 buf_offset += (uint32_t)mapped_len;
1123 async_iter_finalize(async);
1127 async_iter_cancel(async);
1132 static __rte_always_inline int
1133 vhost_enqueue_single_packed(struct virtio_net *dev,
1134 struct vhost_virtqueue *vq,
1135 struct rte_mbuf *pkt,
1136 struct buf_vector *buf_vec,
1139 uint16_t nr_vec = 0;
1140 uint16_t avail_idx = vq->last_avail_idx;
1141 uint16_t max_tries, tries = 0;
1142 uint16_t buf_id = 0;
1144 uint16_t desc_count;
1145 uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1146 uint16_t num_buffers = 0;
1147 uint32_t buffer_len[vq->size];
1148 uint16_t buffer_buf_id[vq->size];
1149 uint16_t buffer_desc_count[vq->size];
1151 if (rxvq_is_mergeable(dev))
1152 max_tries = vq->size - 1;
1158 * if we tried all available ring items, and still
1159 * can't get enough buf, it means something abnormal
1162 if (unlikely(++tries > max_tries))
1165 if (unlikely(fill_vec_buf_packed(dev, vq,
1166 avail_idx, &desc_count,
1169 VHOST_ACCESS_RW) < 0))
1172 len = RTE_MIN(len, size);
1175 buffer_len[num_buffers] = len;
1176 buffer_buf_id[num_buffers] = buf_id;
1177 buffer_desc_count[num_buffers] = desc_count;
1180 *nr_descs += desc_count;
1181 avail_idx += desc_count;
1182 if (avail_idx >= vq->size)
1183 avail_idx -= vq->size;
1186 if (copy_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, num_buffers) < 0)
1189 vhost_shadow_enqueue_single_packed(dev, vq, buffer_len, buffer_buf_id,
1190 buffer_desc_count, num_buffers);
1195 static __rte_noinline uint32_t
1196 virtio_dev_rx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
1197 struct rte_mbuf **pkts, uint32_t count)
1199 uint32_t pkt_idx = 0;
1200 uint16_t num_buffers;
1201 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1202 uint16_t avail_head;
1205 * The ordering between avail index and
1206 * desc reads needs to be enforced.
1208 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1210 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1212 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1213 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1214 uint16_t nr_vec = 0;
1216 if (unlikely(reserve_avail_buf_split(dev, vq,
1217 pkt_len, buf_vec, &num_buffers,
1218 avail_head, &nr_vec) < 0)) {
1219 VHOST_LOG_DATA(DEBUG,
1220 "(%d) failed to get enough desc from vring\n",
1222 vq->shadow_used_idx -= num_buffers;
1226 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1227 dev->vid, vq->last_avail_idx,
1228 vq->last_avail_idx + num_buffers);
1230 if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
1233 vq->shadow_used_idx -= num_buffers;
1237 vq->last_avail_idx += num_buffers;
1240 do_data_copy_enqueue(dev, vq);
1242 if (likely(vq->shadow_used_idx)) {
1243 flush_shadow_used_ring_split(dev, vq);
1244 vhost_vring_call_split(dev, vq);
1250 static __rte_always_inline int
1251 virtio_dev_rx_sync_batch_check(struct virtio_net *dev,
1252 struct vhost_virtqueue *vq,
1253 struct rte_mbuf **pkts,
1254 uint64_t *desc_addrs,
1257 bool wrap_counter = vq->avail_wrap_counter;
1258 struct vring_packed_desc *descs = vq->desc_packed;
1259 uint16_t avail_idx = vq->last_avail_idx;
1260 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1263 if (unlikely(avail_idx & PACKED_BATCH_MASK))
1266 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
1269 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1270 if (unlikely(pkts[i]->next != NULL))
1272 if (unlikely(!desc_is_avail(&descs[avail_idx + i],
1277 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1278 lens[i] = descs[avail_idx + i].len;
1280 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1281 if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
1285 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1286 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
1287 descs[avail_idx + i].addr,
1291 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1292 if (unlikely(!desc_addrs[i]))
1294 if (unlikely(lens[i] != descs[avail_idx + i].len))
1301 static __rte_always_inline void
1302 virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
1303 struct vhost_virtqueue *vq,
1304 struct rte_mbuf **pkts,
1305 uint64_t *desc_addrs,
1308 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1309 struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
1310 struct vring_packed_desc *descs = vq->desc_packed;
1311 uint16_t avail_idx = vq->last_avail_idx;
1312 uint16_t ids[PACKED_BATCH_SIZE];
1315 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1316 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
1317 hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)
1318 (uintptr_t)desc_addrs[i];
1319 lens[i] = pkts[i]->pkt_len +
1320 sizeof(struct virtio_net_hdr_mrg_rxbuf);
1323 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1324 virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
1326 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
1328 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
1329 rte_memcpy((void *)(uintptr_t)(desc_addrs[i] + buf_offset),
1330 rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
1334 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1335 vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr,
1338 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
1339 ids[i] = descs[avail_idx + i].id;
1341 vhost_flush_enqueue_batch_packed(dev, vq, lens, ids);
1344 static __rte_always_inline int
1345 virtio_dev_rx_sync_batch_packed(struct virtio_net *dev,
1346 struct vhost_virtqueue *vq,
1347 struct rte_mbuf **pkts)
1349 uint64_t desc_addrs[PACKED_BATCH_SIZE];
1350 uint64_t lens[PACKED_BATCH_SIZE];
1352 if (virtio_dev_rx_sync_batch_check(dev, vq, pkts, desc_addrs, lens) == -1)
1355 if (vq->shadow_used_idx) {
1356 do_data_copy_enqueue(dev, vq);
1357 vhost_flush_enqueue_shadow_packed(dev, vq);
1360 virtio_dev_rx_batch_packed_copy(dev, vq, pkts, desc_addrs, lens);
1365 static __rte_always_inline int16_t
1366 virtio_dev_rx_single_packed(struct virtio_net *dev,
1367 struct vhost_virtqueue *vq,
1368 struct rte_mbuf *pkt)
1370 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1371 uint16_t nr_descs = 0;
1373 if (unlikely(vhost_enqueue_single_packed(dev, vq, pkt, buf_vec,
1375 VHOST_LOG_DATA(DEBUG,
1376 "(%d) failed to get enough desc from vring\n",
1381 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1382 dev->vid, vq->last_avail_idx,
1383 vq->last_avail_idx + nr_descs);
1385 vq_inc_last_avail_packed(vq, nr_descs);
1390 static __rte_noinline uint32_t
1391 virtio_dev_rx_packed(struct virtio_net *dev,
1392 struct vhost_virtqueue *__rte_restrict vq,
1393 struct rte_mbuf **__rte_restrict pkts,
1396 uint32_t pkt_idx = 0;
1399 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1401 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
1402 if (!virtio_dev_rx_sync_batch_packed(dev, vq,
1404 pkt_idx += PACKED_BATCH_SIZE;
1409 if (virtio_dev_rx_single_packed(dev, vq, pkts[pkt_idx]))
1413 } while (pkt_idx < count);
1415 if (vq->shadow_used_idx) {
1416 do_data_copy_enqueue(dev, vq);
1417 vhost_flush_enqueue_shadow_packed(dev, vq);
1421 vhost_vring_call_packed(dev, vq);
1426 static __rte_always_inline uint32_t
1427 virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
1428 struct rte_mbuf **pkts, uint32_t count)
1430 struct vhost_virtqueue *vq;
1433 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
1434 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
1435 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
1436 dev->vid, __func__, queue_id);
1440 vq = dev->virtqueue[queue_id];
1442 rte_spinlock_lock(&vq->access_lock);
1444 if (unlikely(!vq->enabled))
1445 goto out_access_unlock;
1447 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1448 vhost_user_iotlb_rd_lock(vq);
1450 if (unlikely(!vq->access_ok))
1451 if (unlikely(vring_translate(dev, vq) < 0))
1454 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
1458 if (vq_is_packed(dev))
1459 nb_tx = virtio_dev_rx_packed(dev, vq, pkts, count);
1461 nb_tx = virtio_dev_rx_split(dev, vq, pkts, count);
1464 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
1465 vhost_user_iotlb_rd_unlock(vq);
1468 rte_spinlock_unlock(&vq->access_lock);
1474 rte_vhost_enqueue_burst(int vid, uint16_t queue_id,
1475 struct rte_mbuf **__rte_restrict pkts, uint16_t count)
1477 struct virtio_net *dev = get_device(vid);
1482 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
1484 "(%d) %s: built-in vhost net backend is disabled.\n",
1485 dev->vid, __func__);
1489 return virtio_dev_rx(dev, queue_id, pkts, count);
1492 static __rte_always_inline uint16_t
1493 virtio_dev_rx_async_get_info_idx(uint16_t pkts_idx,
1494 uint16_t vq_size, uint16_t n_inflight)
1496 return pkts_idx > n_inflight ? (pkts_idx - n_inflight) :
1497 (vq_size - n_inflight + pkts_idx) % vq_size;
1500 static __rte_always_inline void
1501 store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem *d_ring,
1502 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1504 size_t elem_size = sizeof(struct vring_used_elem);
1506 if (d_idx + count <= ring_size) {
1507 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1509 uint16_t size = ring_size - d_idx;
1511 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1512 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1516 static __rte_always_inline void
1517 store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
1518 struct vring_used_elem_packed *d_ring,
1519 uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
1521 size_t elem_size = sizeof(struct vring_used_elem_packed);
1523 if (d_idx + count <= ring_size) {
1524 rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
1526 uint16_t size = ring_size - d_idx;
1528 rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
1529 rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
1533 static __rte_noinline uint32_t
1534 virtio_dev_rx_async_submit_split(struct virtio_net *dev,
1535 struct vhost_virtqueue *vq, uint16_t queue_id,
1536 struct rte_mbuf **pkts, uint32_t count)
1538 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1539 uint32_t pkt_idx = 0;
1540 uint16_t num_buffers;
1541 uint16_t avail_head;
1543 struct vhost_async *async = vq->async;
1544 struct async_inflight_info *pkts_info = async->pkts_info;
1545 uint32_t pkt_err = 0;
1547 uint16_t slot_idx = 0;
1550 * The ordering between avail index and desc reads need to be enforced.
1552 avail_head = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE);
1554 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
1556 async_iter_reset(async);
1558 for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
1559 uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
1560 uint16_t nr_vec = 0;
1562 if (unlikely(reserve_avail_buf_split(dev, vq, pkt_len, buf_vec,
1563 &num_buffers, avail_head, &nr_vec) < 0)) {
1564 VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n",
1566 vq->shadow_used_idx -= num_buffers;
1570 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1571 dev->vid, vq->last_avail_idx, vq->last_avail_idx + num_buffers);
1573 if (async_mbuf_to_desc(dev, vq, pkts[pkt_idx], buf_vec, nr_vec, num_buffers) < 0) {
1574 vq->shadow_used_idx -= num_buffers;
1578 slot_idx = (async->pkts_idx + pkt_idx) & (vq->size - 1);
1579 pkts_info[slot_idx].descs = num_buffers;
1580 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1582 vq->last_avail_idx += num_buffers;
1585 if (unlikely(pkt_idx == 0))
1588 n_xfer = async->ops.transfer_data(dev->vid, queue_id, async->iov_iter, 0, pkt_idx);
1589 if (unlikely(n_xfer < 0)) {
1590 VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
1591 dev->vid, __func__, queue_id);
1595 pkt_err = pkt_idx - n_xfer;
1596 if (unlikely(pkt_err)) {
1597 uint16_t num_descs = 0;
1599 /* update number of completed packets */
1602 /* calculate the sum of descriptors to revert */
1603 while (pkt_err-- > 0) {
1604 num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
1608 /* recover shadow used ring and available ring */
1609 vq->shadow_used_idx -= num_descs;
1610 vq->last_avail_idx -= num_descs;
1613 /* keep used descriptors */
1614 if (likely(vq->shadow_used_idx)) {
1615 uint16_t to = async->desc_idx_split & (vq->size - 1);
1617 store_dma_desc_info_split(vq->shadow_used_split,
1618 async->descs_split, vq->size, 0, to,
1619 vq->shadow_used_idx);
1621 async->desc_idx_split += vq->shadow_used_idx;
1622 async->pkts_idx += pkt_idx;
1623 async->pkts_inflight_n += pkt_idx;
1624 vq->shadow_used_idx = 0;
1630 static __rte_always_inline void
1631 vhost_update_used_packed(struct vhost_virtqueue *vq,
1632 struct vring_used_elem_packed *shadow_ring,
1636 uint16_t used_idx = vq->last_used_idx;
1637 uint16_t head_idx = vq->last_used_idx;
1638 uint16_t head_flags = 0;
1643 /* Split loop in two to save memory barriers */
1644 for (i = 0; i < count; i++) {
1645 vq->desc_packed[used_idx].id = shadow_ring[i].id;
1646 vq->desc_packed[used_idx].len = shadow_ring[i].len;
1648 used_idx += shadow_ring[i].count;
1649 if (used_idx >= vq->size)
1650 used_idx -= vq->size;
1653 /* The ordering for storing desc flags needs to be enforced. */
1654 rte_atomic_thread_fence(__ATOMIC_RELEASE);
1656 for (i = 0; i < count; i++) {
1659 if (vq->shadow_used_packed[i].len)
1660 flags = VRING_DESC_F_WRITE;
1664 if (vq->used_wrap_counter) {
1665 flags |= VRING_DESC_F_USED;
1666 flags |= VRING_DESC_F_AVAIL;
1668 flags &= ~VRING_DESC_F_USED;
1669 flags &= ~VRING_DESC_F_AVAIL;
1673 vq->desc_packed[vq->last_used_idx].flags = flags;
1675 head_idx = vq->last_used_idx;
1679 vq_inc_last_used_packed(vq, shadow_ring[i].count);
1682 vq->desc_packed[head_idx].flags = head_flags;
1685 static __rte_always_inline int
1686 vhost_enqueue_async_packed(struct virtio_net *dev,
1687 struct vhost_virtqueue *vq,
1688 struct rte_mbuf *pkt,
1689 struct buf_vector *buf_vec,
1691 uint16_t *nr_buffers)
1693 uint16_t nr_vec = 0;
1694 uint16_t avail_idx = vq->last_avail_idx;
1695 uint16_t max_tries, tries = 0;
1696 uint16_t buf_id = 0;
1698 uint16_t desc_count = 0;
1699 uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
1700 uint32_t buffer_len[vq->size];
1701 uint16_t buffer_buf_id[vq->size];
1702 uint16_t buffer_desc_count[vq->size];
1704 if (rxvq_is_mergeable(dev))
1705 max_tries = vq->size - 1;
1711 * if we tried all available ring items, and still
1712 * can't get enough buf, it means something abnormal
1715 if (unlikely(++tries > max_tries))
1718 if (unlikely(fill_vec_buf_packed(dev, vq,
1719 avail_idx, &desc_count,
1722 VHOST_ACCESS_RW) < 0))
1725 len = RTE_MIN(len, size);
1728 buffer_len[*nr_buffers] = len;
1729 buffer_buf_id[*nr_buffers] = buf_id;
1730 buffer_desc_count[*nr_buffers] = desc_count;
1732 *nr_descs += desc_count;
1733 avail_idx += desc_count;
1734 if (avail_idx >= vq->size)
1735 avail_idx -= vq->size;
1738 if (unlikely(async_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec,
1742 vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
1747 static __rte_always_inline int16_t
1748 virtio_dev_rx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
1749 struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers)
1751 struct buf_vector buf_vec[BUF_VECTOR_MAX];
1753 if (unlikely(vhost_enqueue_async_packed(dev, vq, pkt, buf_vec,
1754 nr_descs, nr_buffers) < 0)) {
1755 VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n", dev->vid);
1759 VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
1760 dev->vid, vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
1765 static __rte_always_inline void
1766 dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
1767 uint32_t nr_err, uint32_t *pkt_idx)
1769 uint16_t descs_err = 0;
1770 uint16_t buffers_err = 0;
1771 struct async_inflight_info *pkts_info = vq->async->pkts_info;
1774 /* calculate the sum of buffers and descs of DMA-error packets. */
1775 while (nr_err-- > 0) {
1776 descs_err += pkts_info[slot_idx % vq->size].descs;
1777 buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
1781 if (vq->last_avail_idx >= descs_err) {
1782 vq->last_avail_idx -= descs_err;
1784 vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
1785 vq->avail_wrap_counter ^= 1;
1788 vq->shadow_used_idx -= buffers_err;
1791 static __rte_noinline uint32_t
1792 virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
1793 struct vhost_virtqueue *vq, uint16_t queue_id,
1794 struct rte_mbuf **pkts, uint32_t count)
1796 uint32_t pkt_idx = 0;
1797 uint32_t remained = count;
1799 uint16_t num_buffers;
1802 struct vhost_async *async = vq->async;
1803 struct async_inflight_info *pkts_info = async->pkts_info;
1804 uint32_t pkt_err = 0;
1805 uint16_t slot_idx = 0;
1808 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
1812 if (unlikely(virtio_dev_rx_async_packed(dev, vq, pkts[pkt_idx],
1813 &num_descs, &num_buffers) < 0))
1816 slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
1818 pkts_info[slot_idx].descs = num_descs;
1819 pkts_info[slot_idx].nr_buffers = num_buffers;
1820 pkts_info[slot_idx].mbuf = pkts[pkt_idx];
1824 vq_inc_last_avail_packed(vq, num_descs);
1825 } while (pkt_idx < count);
1827 if (unlikely(pkt_idx == 0))
1830 n_xfer = async->ops.transfer_data(dev->vid, queue_id, async->iov_iter, 0, pkt_idx);
1831 if (unlikely(n_xfer < 0)) {
1832 VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
1833 dev->vid, __func__, queue_id);
1837 pkt_err = pkt_idx - n_xfer;
1839 async_iter_reset(async);
1841 if (unlikely(pkt_err))
1842 dma_error_handler_packed(vq, slot_idx, pkt_err, &pkt_idx);
1844 if (likely(vq->shadow_used_idx)) {
1845 /* keep used descriptors. */
1846 store_dma_desc_info_packed(vq->shadow_used_packed, async->buffers_packed,
1847 vq->size, 0, async->buffer_idx_packed,
1848 vq->shadow_used_idx);
1850 async->buffer_idx_packed += vq->shadow_used_idx;
1851 if (async->buffer_idx_packed >= vq->size)
1852 async->buffer_idx_packed -= vq->size;
1854 async->pkts_idx += pkt_idx;
1855 if (async->pkts_idx >= vq->size)
1856 async->pkts_idx -= vq->size;
1858 vq->shadow_used_idx = 0;
1859 async->pkts_inflight_n += pkt_idx;
1865 static __rte_always_inline void
1866 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
1868 struct vhost_async *async = vq->async;
1869 uint16_t nr_left = n_descs;
1874 from = async->last_desc_idx_split & (vq->size - 1);
1875 nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
1876 to = vq->last_used_idx & (vq->size - 1);
1878 if (to + nr_copy <= vq->size) {
1879 rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1880 nr_copy * sizeof(struct vring_used_elem));
1882 uint16_t size = vq->size - to;
1884 rte_memcpy(&vq->used->ring[to], &async->descs_split[from],
1885 size * sizeof(struct vring_used_elem));
1886 rte_memcpy(&vq->used->ring[0], &async->descs_split[from + size],
1887 (nr_copy - size) * sizeof(struct vring_used_elem));
1890 async->last_desc_idx_split += nr_copy;
1891 vq->last_used_idx += nr_copy;
1893 } while (nr_left > 0);
1896 static __rte_always_inline void
1897 write_back_completed_descs_packed(struct vhost_virtqueue *vq,
1900 struct vhost_async *async = vq->async;
1901 uint16_t nr_left = n_buffers;
1905 from = async->last_buffer_idx_packed;
1906 to = (from + nr_left) % vq->size;
1908 vhost_update_used_packed(vq, async->buffers_packed + from, to - from);
1909 async->last_buffer_idx_packed += nr_left;
1912 vhost_update_used_packed(vq, async->buffers_packed + from,
1914 async->last_buffer_idx_packed = 0;
1915 nr_left -= vq->size - from;
1917 } while (nr_left > 0);
1920 static __rte_always_inline uint16_t
1921 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
1922 struct rte_mbuf **pkts, uint16_t count)
1924 struct vhost_virtqueue *vq;
1925 struct vhost_async *async;
1926 struct async_inflight_info *pkts_info;
1928 uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0, n_buffers = 0;
1929 uint16_t start_idx, pkts_idx, vq_size;
1932 vq = dev->virtqueue[queue_id];
1934 pkts_idx = async->pkts_idx % vq->size;
1935 pkts_info = async->pkts_info;
1937 start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
1938 vq_size, async->pkts_inflight_n);
1940 if (count > async->last_pkts_n) {
1941 n_cpl = async->ops.check_completed_copies(dev->vid,
1942 queue_id, 0, count - async->last_pkts_n);
1943 if (likely(n_cpl >= 0)) {
1947 "(%d) %s: failed to check completed copies for queue id %d.\n",
1948 dev->vid, __func__, queue_id);
1953 n_pkts_cpl += async->last_pkts_n;
1954 n_pkts_put = RTE_MIN(n_pkts_cpl, count);
1955 if (unlikely(n_pkts_put == 0)) {
1956 async->last_pkts_n = n_pkts_cpl;
1960 if (vq_is_packed(dev)) {
1961 for (i = 0; i < n_pkts_put; i++) {
1962 from = (start_idx + i) % vq_size;
1963 n_buffers += pkts_info[from].nr_buffers;
1964 pkts[i] = pkts_info[from].mbuf;
1967 for (i = 0; i < n_pkts_put; i++) {
1968 from = (start_idx + i) & (vq_size - 1);
1969 n_descs += pkts_info[from].descs;
1970 pkts[i] = pkts_info[from].mbuf;
1973 async->last_pkts_n = n_pkts_cpl - n_pkts_put;
1974 async->pkts_inflight_n -= n_pkts_put;
1976 if (likely(vq->enabled && vq->access_ok)) {
1977 if (vq_is_packed(dev)) {
1978 write_back_completed_descs_packed(vq, n_buffers);
1980 vhost_vring_call_packed(dev, vq);
1982 write_back_completed_descs_split(vq, n_descs);
1984 __atomic_add_fetch(&vq->used->idx, n_descs,
1986 vhost_vring_call_split(dev, vq);
1989 if (vq_is_packed(dev)) {
1990 async->last_buffer_idx_packed += n_buffers;
1991 if (async->last_buffer_idx_packed >= vq->size)
1992 async->last_buffer_idx_packed -= vq->size;
1994 async->last_desc_idx_split += n_descs;
2002 rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
2003 struct rte_mbuf **pkts, uint16_t count)
2005 struct virtio_net *dev = get_device(vid);
2006 struct vhost_virtqueue *vq;
2007 uint16_t n_pkts_cpl = 0;
2012 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2013 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2014 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2015 dev->vid, __func__, queue_id);
2019 vq = dev->virtqueue[queue_id];
2021 if (unlikely(!vq->async)) {
2022 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2023 dev->vid, __func__, queue_id);
2027 rte_spinlock_lock(&vq->access_lock);
2029 n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2031 rte_spinlock_unlock(&vq->access_lock);
2037 rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
2038 struct rte_mbuf **pkts, uint16_t count)
2040 struct virtio_net *dev = get_device(vid);
2041 struct vhost_virtqueue *vq;
2042 uint16_t n_pkts_cpl = 0;
2047 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2048 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2049 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2050 dev->vid, __func__, queue_id);
2054 vq = dev->virtqueue[queue_id];
2056 if (unlikely(!vq->async)) {
2057 VHOST_LOG_DATA(ERR, "(%d) %s: async not registered for queue id %d.\n",
2058 dev->vid, __func__, queue_id);
2062 n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
2067 static __rte_always_inline uint32_t
2068 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
2069 struct rte_mbuf **pkts, uint32_t count)
2071 struct vhost_virtqueue *vq;
2074 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2075 if (unlikely(!is_valid_virt_queue_idx(queue_id, 0, dev->nr_vring))) {
2076 VHOST_LOG_DATA(ERR, "(%d) %s: invalid virtqueue idx %d.\n",
2077 dev->vid, __func__, queue_id);
2081 vq = dev->virtqueue[queue_id];
2083 rte_spinlock_lock(&vq->access_lock);
2085 if (unlikely(!vq->enabled || !vq->async))
2086 goto out_access_unlock;
2088 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2089 vhost_user_iotlb_rd_lock(vq);
2091 if (unlikely(!vq->access_ok))
2092 if (unlikely(vring_translate(dev, vq) < 0))
2095 count = RTE_MIN((uint32_t)MAX_PKT_BURST, count);
2099 if (vq_is_packed(dev))
2100 nb_tx = virtio_dev_rx_async_submit_packed(dev, vq, queue_id,
2103 nb_tx = virtio_dev_rx_async_submit_split(dev, vq, queue_id,
2107 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
2108 vhost_user_iotlb_rd_unlock(vq);
2111 rte_spinlock_unlock(&vq->access_lock);
2117 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
2118 struct rte_mbuf **pkts, uint16_t count)
2120 struct virtio_net *dev = get_device(vid);
2125 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
2127 "(%d) %s: built-in vhost net backend is disabled.\n",
2128 dev->vid, __func__);
2132 return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
2136 virtio_net_with_host_offload(struct virtio_net *dev)
2139 ((1ULL << VIRTIO_NET_F_CSUM) |
2140 (1ULL << VIRTIO_NET_F_HOST_ECN) |
2141 (1ULL << VIRTIO_NET_F_HOST_TSO4) |
2142 (1ULL << VIRTIO_NET_F_HOST_TSO6) |
2143 (1ULL << VIRTIO_NET_F_HOST_UFO)))
2150 parse_headers(struct rte_mbuf *m, uint8_t *l4_proto)
2152 struct rte_ipv4_hdr *ipv4_hdr;
2153 struct rte_ipv6_hdr *ipv6_hdr;
2154 struct rte_ether_hdr *eth_hdr;
2156 uint16_t data_len = rte_pktmbuf_data_len(m);
2158 if (data_len < sizeof(struct rte_ether_hdr))
2161 eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
2163 m->l2_len = sizeof(struct rte_ether_hdr);
2164 ethertype = rte_be_to_cpu_16(eth_hdr->ether_type);
2166 if (ethertype == RTE_ETHER_TYPE_VLAN) {
2167 if (data_len < sizeof(struct rte_ether_hdr) +
2168 sizeof(struct rte_vlan_hdr))
2171 struct rte_vlan_hdr *vlan_hdr =
2172 (struct rte_vlan_hdr *)(eth_hdr + 1);
2174 m->l2_len += sizeof(struct rte_vlan_hdr);
2175 ethertype = rte_be_to_cpu_16(vlan_hdr->eth_proto);
2178 switch (ethertype) {
2179 case RTE_ETHER_TYPE_IPV4:
2180 if (data_len < m->l2_len + sizeof(struct rte_ipv4_hdr))
2182 ipv4_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
2184 m->l3_len = rte_ipv4_hdr_len(ipv4_hdr);
2185 if (data_len < m->l2_len + m->l3_len)
2187 m->ol_flags |= RTE_MBUF_F_TX_IPV4;
2188 *l4_proto = ipv4_hdr->next_proto_id;
2190 case RTE_ETHER_TYPE_IPV6:
2191 if (data_len < m->l2_len + sizeof(struct rte_ipv6_hdr))
2193 ipv6_hdr = rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
2195 m->l3_len = sizeof(struct rte_ipv6_hdr);
2196 m->ol_flags |= RTE_MBUF_F_TX_IPV6;
2197 *l4_proto = ipv6_hdr->proto;
2200 /* a valid L3 header is needed for further L4 parsing */
2204 /* both CSUM and GSO need a valid L4 header */
2205 switch (*l4_proto) {
2207 if (data_len < m->l2_len + m->l3_len +
2208 sizeof(struct rte_tcp_hdr))
2212 if (data_len < m->l2_len + m->l3_len +
2213 sizeof(struct rte_udp_hdr))
2217 if (data_len < m->l2_len + m->l3_len +
2218 sizeof(struct rte_sctp_hdr))
2234 static __rte_always_inline void
2235 vhost_dequeue_offload_legacy(struct virtio_net_hdr *hdr, struct rte_mbuf *m)
2237 uint8_t l4_proto = 0;
2238 struct rte_tcp_hdr *tcp_hdr = NULL;
2240 uint16_t data_len = rte_pktmbuf_data_len(m);
2242 if (parse_headers(m, &l4_proto) < 0)
2245 if (hdr->flags == VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2246 if (hdr->csum_start == (m->l2_len + m->l3_len)) {
2247 switch (hdr->csum_offset) {
2248 case (offsetof(struct rte_tcp_hdr, cksum)):
2249 if (l4_proto != IPPROTO_TCP)
2251 m->ol_flags |= RTE_MBUF_F_TX_TCP_CKSUM;
2253 case (offsetof(struct rte_udp_hdr, dgram_cksum)):
2254 if (l4_proto != IPPROTO_UDP)
2256 m->ol_flags |= RTE_MBUF_F_TX_UDP_CKSUM;
2258 case (offsetof(struct rte_sctp_hdr, cksum)):
2259 if (l4_proto != IPPROTO_SCTP)
2261 m->ol_flags |= RTE_MBUF_F_TX_SCTP_CKSUM;
2271 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2272 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2273 case VIRTIO_NET_HDR_GSO_TCPV4:
2274 case VIRTIO_NET_HDR_GSO_TCPV6:
2275 if (l4_proto != IPPROTO_TCP)
2277 tcp_hdr = rte_pktmbuf_mtod_offset(m,
2278 struct rte_tcp_hdr *,
2279 m->l2_len + m->l3_len);
2280 tcp_len = (tcp_hdr->data_off & 0xf0) >> 2;
2281 if (data_len < m->l2_len + m->l3_len + tcp_len)
2283 m->ol_flags |= RTE_MBUF_F_TX_TCP_SEG;
2284 m->tso_segsz = hdr->gso_size;
2285 m->l4_len = tcp_len;
2287 case VIRTIO_NET_HDR_GSO_UDP:
2288 if (l4_proto != IPPROTO_UDP)
2290 m->ol_flags |= RTE_MBUF_F_TX_UDP_SEG;
2291 m->tso_segsz = hdr->gso_size;
2292 m->l4_len = sizeof(struct rte_udp_hdr);
2295 VHOST_LOG_DATA(WARNING,
2296 "unsupported gso type %u.\n", hdr->gso_type);
2308 static __rte_always_inline void
2309 vhost_dequeue_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *m,
2310 bool legacy_ol_flags)
2312 struct rte_net_hdr_lens hdr_lens;
2313 int l4_supported = 0;
2316 if (hdr->flags == 0 && hdr->gso_type == VIRTIO_NET_HDR_GSO_NONE)
2319 if (legacy_ol_flags) {
2320 vhost_dequeue_offload_legacy(hdr, m);
2324 m->ol_flags |= RTE_MBUF_F_RX_IP_CKSUM_UNKNOWN;
2326 ptype = rte_net_get_ptype(m, &hdr_lens, RTE_PTYPE_ALL_MASK);
2327 m->packet_type = ptype;
2328 if ((ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_TCP ||
2329 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_UDP ||
2330 (ptype & RTE_PTYPE_L4_MASK) == RTE_PTYPE_L4_SCTP)
2333 /* According to Virtio 1.1 spec, the device only needs to look at
2334 * VIRTIO_NET_HDR_F_NEEDS_CSUM in the packet transmission path.
2335 * This differs from the processing incoming packets path where the
2336 * driver could rely on VIRTIO_NET_HDR_F_DATA_VALID flag set by the
2339 * 5.1.6.2.1 Driver Requirements: Packet Transmission
2340 * The driver MUST NOT set the VIRTIO_NET_HDR_F_DATA_VALID and
2341 * VIRTIO_NET_HDR_F_RSC_INFO bits in flags.
2343 * 5.1.6.2.2 Device Requirements: Packet Transmission
2344 * The device MUST ignore flag bits that it does not recognize.
2346 if (hdr->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
2349 hdrlen = hdr_lens.l2_len + hdr_lens.l3_len + hdr_lens.l4_len;
2350 if (hdr->csum_start <= hdrlen && l4_supported != 0) {
2351 m->ol_flags |= RTE_MBUF_F_RX_L4_CKSUM_NONE;
2353 /* Unknown proto or tunnel, do sw cksum. We can assume
2354 * the cksum field is in the first segment since the
2355 * buffers we provided to the host are large enough.
2356 * In case of SCTP, this will be wrong since it's a CRC
2357 * but there's nothing we can do.
2359 uint16_t csum = 0, off;
2361 if (rte_raw_cksum_mbuf(m, hdr->csum_start,
2362 rte_pktmbuf_pkt_len(m) - hdr->csum_start, &csum) < 0)
2364 if (likely(csum != 0xffff))
2366 off = hdr->csum_offset + hdr->csum_start;
2367 if (rte_pktmbuf_data_len(m) >= off + 1)
2368 *rte_pktmbuf_mtod_offset(m, uint16_t *, off) = csum;
2372 if (hdr->gso_type != VIRTIO_NET_HDR_GSO_NONE) {
2373 if (hdr->gso_size == 0)
2376 switch (hdr->gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
2377 case VIRTIO_NET_HDR_GSO_TCPV4:
2378 case VIRTIO_NET_HDR_GSO_TCPV6:
2379 if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_TCP)
2381 m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2382 m->tso_segsz = hdr->gso_size;
2384 case VIRTIO_NET_HDR_GSO_UDP:
2385 if ((ptype & RTE_PTYPE_L4_MASK) != RTE_PTYPE_L4_UDP)
2387 m->ol_flags |= RTE_MBUF_F_RX_LRO | RTE_MBUF_F_RX_L4_CKSUM_NONE;
2388 m->tso_segsz = hdr->gso_size;
2396 static __rte_noinline void
2397 copy_vnet_hdr_from_desc(struct virtio_net_hdr *hdr,
2398 struct buf_vector *buf_vec)
2401 uint64_t remain = sizeof(struct virtio_net_hdr);
2403 uint64_t dst = (uint64_t)(uintptr_t)hdr;
2406 len = RTE_MIN(remain, buf_vec->buf_len);
2407 src = buf_vec->buf_addr;
2408 rte_memcpy((void *)(uintptr_t)dst,
2409 (void *)(uintptr_t)src, len);
2417 static __rte_always_inline int
2418 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
2419 struct buf_vector *buf_vec, uint16_t nr_vec,
2420 struct rte_mbuf *m, struct rte_mempool *mbuf_pool,
2421 bool legacy_ol_flags)
2423 uint32_t buf_avail, buf_offset;
2424 uint64_t buf_addr, buf_len;
2425 uint32_t mbuf_avail, mbuf_offset;
2427 struct rte_mbuf *cur = m, *prev = m;
2428 struct virtio_net_hdr tmp_hdr;
2429 struct virtio_net_hdr *hdr = NULL;
2430 /* A counter to avoid desc dead loop chain */
2431 uint16_t vec_idx = 0;
2432 struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
2435 buf_addr = buf_vec[vec_idx].buf_addr;
2436 buf_len = buf_vec[vec_idx].buf_len;
2438 if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
2443 if (virtio_net_with_host_offload(dev)) {
2444 if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
2446 * No luck, the virtio-net header doesn't fit
2447 * in a contiguous virtual area.
2449 copy_vnet_hdr_from_desc(&tmp_hdr, buf_vec);
2452 hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
2457 * A virtio driver normally uses at least 2 desc buffers
2458 * for Tx: the first for storing the header, and others
2459 * for storing the data.
2461 if (unlikely(buf_len < dev->vhost_hlen)) {
2462 buf_offset = dev->vhost_hlen - buf_len;
2464 buf_addr = buf_vec[vec_idx].buf_addr;
2465 buf_len = buf_vec[vec_idx].buf_len;
2466 buf_avail = buf_len - buf_offset;
2467 } else if (buf_len == dev->vhost_hlen) {
2468 if (unlikely(++vec_idx >= nr_vec))
2470 buf_addr = buf_vec[vec_idx].buf_addr;
2471 buf_len = buf_vec[vec_idx].buf_len;
2474 buf_avail = buf_len;
2476 buf_offset = dev->vhost_hlen;
2477 buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
2481 (uintptr_t)(buf_addr + buf_offset),
2482 (uint32_t)buf_avail, 0);
2485 mbuf_avail = m->buf_len - RTE_PKTMBUF_HEADROOM;
2487 cpy_len = RTE_MIN(buf_avail, mbuf_avail);
2489 if (likely(cpy_len > MAX_BATCH_LEN ||
2490 vq->batch_copy_nb_elems >= vq->size ||
2491 (hdr && cur == m))) {
2492 rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
2494 (void *)((uintptr_t)(buf_addr +
2495 buf_offset)), cpy_len);
2497 batch_copy[vq->batch_copy_nb_elems].dst =
2498 rte_pktmbuf_mtod_offset(cur, void *,
2500 batch_copy[vq->batch_copy_nb_elems].src =
2501 (void *)((uintptr_t)(buf_addr + buf_offset));
2502 batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
2503 vq->batch_copy_nb_elems++;
2506 mbuf_avail -= cpy_len;
2507 mbuf_offset += cpy_len;
2508 buf_avail -= cpy_len;
2509 buf_offset += cpy_len;
2511 /* This buf reaches to its end, get the next one */
2512 if (buf_avail == 0) {
2513 if (++vec_idx >= nr_vec)
2516 buf_addr = buf_vec[vec_idx].buf_addr;
2517 buf_len = buf_vec[vec_idx].buf_len;
2520 buf_avail = buf_len;
2522 PRINT_PACKET(dev, (uintptr_t)buf_addr,
2523 (uint32_t)buf_avail, 0);
2527 * This mbuf reaches to its end, get a new one
2528 * to hold more data.
2530 if (mbuf_avail == 0) {
2531 cur = rte_pktmbuf_alloc(mbuf_pool);
2532 if (unlikely(cur == NULL)) {
2533 VHOST_LOG_DATA(ERR, "Failed to "
2534 "allocate memory for mbuf.\n");
2540 prev->data_len = mbuf_offset;
2542 m->pkt_len += mbuf_offset;
2546 mbuf_avail = cur->buf_len - RTE_PKTMBUF_HEADROOM;
2550 prev->data_len = mbuf_offset;
2551 m->pkt_len += mbuf_offset;
2554 vhost_dequeue_offload(hdr, m, legacy_ol_flags);
2562 virtio_dev_extbuf_free(void *addr __rte_unused, void *opaque)
2568 virtio_dev_extbuf_alloc(struct rte_mbuf *pkt, uint32_t size)
2570 struct rte_mbuf_ext_shared_info *shinfo = NULL;
2571 uint32_t total_len = RTE_PKTMBUF_HEADROOM + size;
2576 total_len += sizeof(*shinfo) + sizeof(uintptr_t);
2577 total_len = RTE_ALIGN_CEIL(total_len, sizeof(uintptr_t));
2579 if (unlikely(total_len > UINT16_MAX))
2582 buf_len = total_len;
2583 buf = rte_malloc(NULL, buf_len, RTE_CACHE_LINE_SIZE);
2584 if (unlikely(buf == NULL))
2587 /* Initialize shinfo */
2588 shinfo = rte_pktmbuf_ext_shinfo_init_helper(buf, &buf_len,
2589 virtio_dev_extbuf_free, buf);
2590 if (unlikely(shinfo == NULL)) {
2592 VHOST_LOG_DATA(ERR, "Failed to init shinfo\n");
2596 iova = rte_malloc_virt2iova(buf);
2597 rte_pktmbuf_attach_extbuf(pkt, buf, iova, buf_len, shinfo);
2598 rte_pktmbuf_reset_headroom(pkt);
2604 * Prepare a host supported pktmbuf.
2606 static __rte_always_inline int
2607 virtio_dev_pktmbuf_prep(struct virtio_net *dev, struct rte_mbuf *pkt,
2610 if (rte_pktmbuf_tailroom(pkt) >= data_len)
2613 /* attach an external buffer if supported */
2614 if (dev->extbuf && !virtio_dev_extbuf_alloc(pkt, data_len))
2617 /* check if chained buffers are allowed */
2618 if (!dev->linearbuf)
2626 virtio_dev_tx_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
2627 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count,
2628 bool legacy_ol_flags)
2631 uint16_t free_entries;
2632 uint16_t dropped = 0;
2633 static bool allocerr_warned;
2636 * The ordering between avail index and
2637 * desc reads needs to be enforced.
2639 free_entries = __atomic_load_n(&vq->avail->idx, __ATOMIC_ACQUIRE) -
2641 if (free_entries == 0)
2644 rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
2646 VHOST_LOG_DATA(DEBUG, "(%d) %s\n", dev->vid, __func__);
2648 count = RTE_MIN(count, MAX_PKT_BURST);
2649 count = RTE_MIN(count, free_entries);
2650 VHOST_LOG_DATA(DEBUG, "(%d) about to dequeue %u buffers\n",
2653 if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2656 for (i = 0; i < count; i++) {
2657 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2660 uint16_t nr_vec = 0;
2663 if (unlikely(fill_vec_buf_split(dev, vq,
2664 vq->last_avail_idx + i,
2666 &head_idx, &buf_len,
2667 VHOST_ACCESS_RO) < 0))
2670 update_shadow_used_ring_split(vq, head_idx, 0);
2672 err = virtio_dev_pktmbuf_prep(dev, pkts[i], buf_len);
2673 if (unlikely(err)) {
2675 * mbuf allocation fails for jumbo packets when external
2676 * buffer allocation is not allowed and linear buffer
2677 * is required. Drop this packet.
2679 if (!allocerr_warned) {
2681 "Failed mbuf alloc of size %d from %s on %s.\n",
2682 buf_len, mbuf_pool->name, dev->ifname);
2683 allocerr_warned = true;
2690 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
2691 mbuf_pool, legacy_ol_flags);
2692 if (unlikely(err)) {
2693 if (!allocerr_warned) {
2695 "Failed to copy desc to mbuf on %s.\n",
2697 allocerr_warned = true;
2706 rte_pktmbuf_free_bulk(&pkts[i - 1], count - i + 1);
2708 vq->last_avail_idx += i;
2710 do_data_copy_dequeue(vq);
2711 if (unlikely(i < count))
2712 vq->shadow_used_idx = i;
2713 if (likely(vq->shadow_used_idx)) {
2714 flush_shadow_used_ring_split(dev, vq);
2715 vhost_vring_call_split(dev, vq);
2718 return (i - dropped);
2723 virtio_dev_tx_split_legacy(struct virtio_net *dev,
2724 struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2725 struct rte_mbuf **pkts, uint16_t count)
2727 return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, true);
2732 virtio_dev_tx_split_compliant(struct virtio_net *dev,
2733 struct vhost_virtqueue *vq, struct rte_mempool *mbuf_pool,
2734 struct rte_mbuf **pkts, uint16_t count)
2736 return virtio_dev_tx_split(dev, vq, mbuf_pool, pkts, count, false);
2739 static __rte_always_inline int
2740 vhost_reserve_avail_batch_packed(struct virtio_net *dev,
2741 struct vhost_virtqueue *vq,
2742 struct rte_mbuf **pkts,
2744 uintptr_t *desc_addrs,
2747 bool wrap = vq->avail_wrap_counter;
2748 struct vring_packed_desc *descs = vq->desc_packed;
2749 uint64_t lens[PACKED_BATCH_SIZE];
2750 uint64_t buf_lens[PACKED_BATCH_SIZE];
2751 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2754 if (unlikely(avail_idx & PACKED_BATCH_MASK))
2756 if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
2759 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2760 flags = descs[avail_idx + i].flags;
2761 if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
2762 (wrap == !!(flags & VRING_DESC_F_USED)) ||
2763 (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
2767 rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
2769 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2770 lens[i] = descs[avail_idx + i].len;
2772 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2773 desc_addrs[i] = vhost_iova_to_vva(dev, vq,
2774 descs[avail_idx + i].addr,
2775 &lens[i], VHOST_ACCESS_RW);
2778 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2779 if (unlikely(!desc_addrs[i]))
2781 if (unlikely((lens[i] != descs[avail_idx + i].len)))
2785 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2786 if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
2790 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2791 buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
2793 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2794 if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
2798 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2799 pkts[i]->pkt_len = lens[i] - buf_offset;
2800 pkts[i]->data_len = pkts[i]->pkt_len;
2801 ids[i] = descs[avail_idx + i].id;
2810 static __rte_always_inline int
2811 virtio_dev_tx_batch_packed(struct virtio_net *dev,
2812 struct vhost_virtqueue *vq,
2813 struct rte_mbuf **pkts,
2814 bool legacy_ol_flags)
2816 uint16_t avail_idx = vq->last_avail_idx;
2817 uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2818 struct virtio_net_hdr *hdr;
2819 uintptr_t desc_addrs[PACKED_BATCH_SIZE];
2820 uint16_t ids[PACKED_BATCH_SIZE];
2823 if (vhost_reserve_avail_batch_packed(dev, vq, pkts, avail_idx,
2827 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2828 rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
2830 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
2831 rte_memcpy(rte_pktmbuf_mtod_offset(pkts[i], void *, 0),
2832 (void *)(uintptr_t)(desc_addrs[i] + buf_offset),
2835 if (virtio_net_with_host_offload(dev)) {
2836 vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
2837 hdr = (struct virtio_net_hdr *)(desc_addrs[i]);
2838 vhost_dequeue_offload(hdr, pkts[i], legacy_ol_flags);
2842 if (virtio_net_is_inorder(dev))
2843 vhost_shadow_dequeue_batch_packed_inorder(vq,
2844 ids[PACKED_BATCH_SIZE - 1]);
2846 vhost_shadow_dequeue_batch_packed(dev, vq, ids);
2848 vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
2853 static __rte_always_inline int
2854 vhost_dequeue_single_packed(struct virtio_net *dev,
2855 struct vhost_virtqueue *vq,
2856 struct rte_mempool *mbuf_pool,
2857 struct rte_mbuf *pkts,
2859 uint16_t *desc_count,
2860 bool legacy_ol_flags)
2862 struct buf_vector buf_vec[BUF_VECTOR_MAX];
2864 uint16_t nr_vec = 0;
2866 static bool allocerr_warned;
2868 if (unlikely(fill_vec_buf_packed(dev, vq,
2869 vq->last_avail_idx, desc_count,
2872 VHOST_ACCESS_RO) < 0))
2875 if (unlikely(virtio_dev_pktmbuf_prep(dev, pkts, buf_len))) {
2876 if (!allocerr_warned) {
2878 "Failed mbuf alloc of size %d from %s on %s.\n",
2879 buf_len, mbuf_pool->name, dev->ifname);
2880 allocerr_warned = true;
2885 err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts,
2886 mbuf_pool, legacy_ol_flags);
2887 if (unlikely(err)) {
2888 if (!allocerr_warned) {
2890 "Failed to copy desc to mbuf on %s.\n",
2892 allocerr_warned = true;
2900 static __rte_always_inline int
2901 virtio_dev_tx_single_packed(struct virtio_net *dev,
2902 struct vhost_virtqueue *vq,
2903 struct rte_mempool *mbuf_pool,
2904 struct rte_mbuf *pkts,
2905 bool legacy_ol_flags)
2908 uint16_t buf_id, desc_count = 0;
2911 ret = vhost_dequeue_single_packed(dev, vq, mbuf_pool, pkts, &buf_id,
2912 &desc_count, legacy_ol_flags);
2914 if (likely(desc_count > 0)) {
2915 if (virtio_net_is_inorder(dev))
2916 vhost_shadow_dequeue_single_packed_inorder(vq, buf_id,
2919 vhost_shadow_dequeue_single_packed(vq, buf_id,
2922 vq_inc_last_avail_packed(vq, desc_count);
2930 virtio_dev_tx_packed(struct virtio_net *dev,
2931 struct vhost_virtqueue *__rte_restrict vq,
2932 struct rte_mempool *mbuf_pool,
2933 struct rte_mbuf **__rte_restrict pkts,
2935 bool legacy_ol_flags)
2937 uint32_t pkt_idx = 0;
2939 if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts, count))
2943 rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
2945 if (count - pkt_idx >= PACKED_BATCH_SIZE) {
2946 if (!virtio_dev_tx_batch_packed(dev, vq,
2949 pkt_idx += PACKED_BATCH_SIZE;
2954 if (virtio_dev_tx_single_packed(dev, vq, mbuf_pool,
2959 } while (pkt_idx < count);
2961 if (pkt_idx != count)
2962 rte_pktmbuf_free_bulk(&pkts[pkt_idx], count - pkt_idx);
2964 if (vq->shadow_used_idx) {
2965 do_data_copy_dequeue(vq);
2967 vhost_flush_dequeue_shadow_packed(dev, vq);
2968 vhost_vring_call_packed(dev, vq);
2976 virtio_dev_tx_packed_legacy(struct virtio_net *dev,
2977 struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
2978 struct rte_mbuf **__rte_restrict pkts, uint32_t count)
2980 return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, true);
2985 virtio_dev_tx_packed_compliant(struct virtio_net *dev,
2986 struct vhost_virtqueue *__rte_restrict vq, struct rte_mempool *mbuf_pool,
2987 struct rte_mbuf **__rte_restrict pkts, uint32_t count)
2989 return virtio_dev_tx_packed(dev, vq, mbuf_pool, pkts, count, false);
2993 rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
2994 struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts, uint16_t count)
2996 struct virtio_net *dev;
2997 struct rte_mbuf *rarp_mbuf = NULL;
2998 struct vhost_virtqueue *vq;
2999 int16_t success = 1;
3001 dev = get_device(vid);
3005 if (unlikely(!(dev->flags & VIRTIO_DEV_BUILTIN_VIRTIO_NET))) {
3007 "(%d) %s: built-in vhost net backend is disabled.\n",
3008 dev->vid, __func__);
3012 if (unlikely(!is_valid_virt_queue_idx(queue_id, 1, dev->nr_vring))) {
3014 "(%d) %s: invalid virtqueue idx %d.\n",
3015 dev->vid, __func__, queue_id);
3019 vq = dev->virtqueue[queue_id];
3021 if (unlikely(rte_spinlock_trylock(&vq->access_lock) == 0))
3024 if (unlikely(!vq->enabled)) {
3026 goto out_access_unlock;
3029 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3030 vhost_user_iotlb_rd_lock(vq);
3032 if (unlikely(!vq->access_ok))
3033 if (unlikely(vring_translate(dev, vq) < 0)) {
3039 * Construct a RARP broadcast packet, and inject it to the "pkts"
3040 * array, to looks like that guest actually send such packet.
3042 * Check user_send_rarp() for more information.
3044 * broadcast_rarp shares a cacheline in the virtio_net structure
3045 * with some fields that are accessed during enqueue and
3046 * __atomic_compare_exchange_n causes a write if performed compare
3047 * and exchange. This could result in false sharing between enqueue
3050 * Prevent unnecessary false sharing by reading broadcast_rarp first
3051 * and only performing compare and exchange if the read indicates it
3052 * is likely to be set.
3054 if (unlikely(__atomic_load_n(&dev->broadcast_rarp, __ATOMIC_ACQUIRE) &&
3055 __atomic_compare_exchange_n(&dev->broadcast_rarp,
3056 &success, 0, 0, __ATOMIC_RELEASE, __ATOMIC_RELAXED))) {
3058 rarp_mbuf = rte_net_make_rarp_packet(mbuf_pool, &dev->mac);
3059 if (rarp_mbuf == NULL) {
3060 VHOST_LOG_DATA(ERR, "Failed to make RARP packet.\n");
3065 * Inject it to the head of "pkts" array, so that switch's mac
3066 * learning table will get updated first.
3068 pkts[0] = rarp_mbuf;
3073 if (vq_is_packed(dev)) {
3074 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3075 count = virtio_dev_tx_packed_legacy(dev, vq, mbuf_pool, pkts, count);
3077 count = virtio_dev_tx_packed_compliant(dev, vq, mbuf_pool, pkts, count);
3079 if (dev->flags & VIRTIO_DEV_LEGACY_OL_FLAGS)
3080 count = virtio_dev_tx_split_legacy(dev, vq, mbuf_pool, pkts, count);
3082 count = virtio_dev_tx_split_compliant(dev, vq, mbuf_pool, pkts, count);
3086 if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
3087 vhost_user_iotlb_rd_unlock(vq);
3090 rte_spinlock_unlock(&vq->access_lock);
3092 if (unlikely(rarp_mbuf != NULL))